Analisis Komparatif Arsitektur CNN dan VGG16 pada Klasifikasi Genre Musik

Abstract

Music genre classification based on spectrogram images is an important task in music information retrieval. This study compares the performance of a custom Convolutional Neural Network (CNN) architecture and VGG-16 for classifying five music genres from the GTZAN dataset: blues, classical, hiphop, metal, and reggae. A total of 500 audio files were converted into spectrogram images for training and testing. The custom CNN was designed and trained from scratch, while VGG-16 utilized pretrained weights with fine-tuning applied to the fully connected layers. Experimental results show that the custom CNN achieved 75% test accuracy and a macro F1- score of 0.74, outperforming VGG-16 which achieved 68.75% accuracy and a macro F1-score of 0.67. These findings demonstrate the advantage of using a tailored architecture for spectrogram- based music genre classification and provide directions for future research, including full fine- tuning of pretrained models, hybrid architectures, and integration of temporal features.